[Pfannkuchensack]

Summary

Stacked on #10040. Review and merge that PR first; this diff includes its commits until it lands.

#10040 added opt-in Tiny AutoEncoder (TAE) live previews for the image families that madebyollin's TAESD-style decoders cover. This PR extends them to the families whose VAE is the Wan or LTX video VAE, using madebyollin's TAEHV decoders (MIT, pinned to commit 62f7591). With tiny_autoencoder_previews: true and the matching starter installed:

Starter	Base	Serves
TAEW2.1	Wan, 16 ch	Wan 2.2 A14B, Qwen-Image (incl. Edit), Krea-2, Anima
TAEW2.2	Wan, 48 ch	Wan 2.2 TI2V-5B
TAELTX2.3	LTX-2	LTX-2.5

How it works:

• TAEHV joins the tiny_autoencoder model type through a decoder base class shared with TAESD. MiniMax H3's built-in taeh3 decoder now builds on it, with its behaviour pinned bit-for-bit by a reference test.
• A single latent decodes as a one-frame clip: the causal warm-up frames are trimmed and the one remaining frame is kept.
• Wan video previews pass a 2-frame window (the middle latent frame plus the one before it). The TAE uses the earlier frame as temporal context. The linear preview still shows the same middle frame as before, pixel-identical.
• Preview caps are per decoder: 512 px for TAEW2.1, whose decode is the heaviest, and 1024 px for all others.
• Some TAE networks share their shapes with decoders for other models, so identification needs a name or an explicit base:
  • taeltx_2 has the same shapes as taeltx2_3 but previews LTX-2.5 worse than linear. It is refused, even with a base override.
  • taew2_1 has the same shapes as the HunyuanVideo (taehv) and CogVideoX (taecvx) decoders. Files named taehv/taecvx are refused.
  • An unnamed file with one of these networks needs the TAE's name or an explicit base. Starters always set the base.
• When no TAE is installed, the log line now names the starter to install and the base it is listed under. For example, TAEW2.1 is listed under Wan.

CogView4 and Ideogram 4 keep linear previews. MiniMax H3 keeps its own decoder.

Related Issues / Discussions

Follows #10040.

QA Instructions

Automated:

• uv run --no-sync pytest -n 4 over tests/backend/tiny_autoencoder, tests/backend/model_manager, tests/backend/architectures, tests/backend/util, tests/model_identification, tests/backend/minimax_h3, tests/app/util, tests/app/services/progress_previews, tests/app/services/session_processor, tests/app/invocations, the preview router and socket tests, tests/test_config.py and tests/test_docs_json_export.py: 7377 passed.
• ruff check and ruff format --check are clean.
• API package: prettier and tests pass. Frontend: tsc passes.
• Docs: pnpm test and pnpm build pass, including link validation.
• The vendored decoder output was compared to upstream taehv.py on identical weights and is bit-equal.

Measured on an RTX 4090, fp16. PSNR is computed against the real VAE decode of the same latent, using the tensor each family's step callback receives:

Family	TAE	TAE	Linear (upscaled)
Qwen-Image / Krea-2	TAEW2.1	41.6 dB	26.6 dB
Anima / Wan image	TAEW2.1	41.2 dB	26.6 dB
Wan 2.2 TI2V-5B	TAEW2.2	34.4 dB	26.3 dB
LTX-2.5	TAELTX2.3	37.8 dB	26.6 dB
LTX-2.5	taeltx_2 (refused)	11.0 dB	26.6 dB
Wan video, middle frame alone / with previous frame	TAEW2.1	20.3 / 34.9 dB	18.3 dB
Wan 5B video, alone / with previous frame	TAEW2.2	19.3 / 27.5 dB	18.1 dB
LTX-2.5 video, middle frame	TAELTX2.3	36.6 dB	27.6 dB

Cost per preview at the caps:

TAE	Cap	VRAM	Time
TAEW2.1	512 px	~0.54 GB	10–22 ms
TAEW2.1 uncapped, 1024² (for comparison)	—	2.1 GB	63 ms
TAEW2.2	1024 px	~0.54 GB	~24 ms
TAELTX2.3	1024 px	~0.27 GB	~29 ms
Wan video 2-frame window	832×480 – 1280×720	~0.6 GB	≤ 16 ms
Wan video 2-frame window	1024²	~1.06 GB	≤ 32 ms

Browser:

• An Anima generation previews through TAEW2.1 at 512 px, with no warnings.
• TAEW2.1 is not offered in Anima's VAE picker.
• Installing taeltx_2 is refused with a message pointing to TAELTX2.3.

Review

Resolved:

• Unnamed LTX decoders and taehv/taecvx files are no longer accepted silently.
• Window downsampling folds frames into the batch and uses 2-D pooling, so it avoids 3-D pooling kernels that may be missing on some backends.
• The window's doubled memory is measured and documented.
• The MiniMax H3 docstring now gives the real reasons it keeps its own path.
• The Wan preview window and the TAELTX2.3 cap are covered by tests.
• The starter lookup stays out of the model configs.

Known limits:

• Qwen-Image, Krea-2 and Wan pass the noisy latent to the step callback, so their TAE previews show sharp noise in early steps. Anima and LTX-2.5 preview the predicted image.
• For very short TI2V-5B image-to-video clips, the context frame can be the conditioned first frame, which drifts during sampling. This affects only the preview.

Compatibility / Rollout

• No new model type and no DB migration. New config classes join the existing tiny_autoencoder type, and openapi.json and schema.ts are regenerated.
• MiniMax H3's decoder module keeps its public names and output.

Checklist

• The PR has a short but descriptive title, suitable for a changelog
• Meaningful regression coverage added / updated where needed; obsolete tests/code removed
• Persisted-state and API changes include required migrations / compatibility validation
• Relevant performance/efficiency opportunities considered; material claims have evidence
• Material review findings resolved and relevant checks rerun
• Documentation added / updated (if applicable)
• Updated What's New copy (if doing a release after this PR)

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}